Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into imperative_mnist

revert-15207-remove_op_handle_lock_and_fix_var
minqiyang 6 years ago
commit 09e2e66236

@ -139,10 +139,12 @@ endfunction()
message(STATUS "CUDA detected: " ${CUDA_VERSION})
if (${CUDA_VERSION} LESS 7.0)
set(paddle_known_gpu_archs ${paddle_known_gpu_archs})
add_definitions("-DPADDLE_CUDA_BINVER=\"60\"")
elseif (${CUDA_VERSION} LESS 8.0) # CUDA 7.x
set(paddle_known_gpu_archs ${paddle_known_gpu_archs7})
list(APPEND CUDA_NVCC_FLAGS "-D_MWAITXINTRIN_H_INCLUDED")
list(APPEND CUDA_NVCC_FLAGS "-D__STRICT_ANSI__")
add_definitions("-DPADDLE_CUDA_BINVER=\"70\"")
elseif (${CUDA_VERSION} LESS 9.0) # CUDA 8.x
set(paddle_known_gpu_archs ${paddle_known_gpu_archs8})
list(APPEND CUDA_NVCC_FLAGS "-D_MWAITXINTRIN_H_INCLUDED")
@ -150,6 +152,7 @@ elseif (${CUDA_VERSION} LESS 9.0) # CUDA 8.x
# CUDA 8 may complain that sm_20 is no longer supported. Suppress the
# warning for now.
list(APPEND CUDA_NVCC_FLAGS "-Wno-deprecated-gpu-targets")
add_definitions("-DPADDLE_CUDA_BINVER=\"80\"")
endif()
include_directories(${CUDA_INCLUDE_DIRS})

@ -89,6 +89,7 @@ if(CUDNN_FOUND)
if(NOT CUDNN_MAJOR_VERSION)
set(CUDNN_VERSION "???")
else()
add_definitions("-DPADDLE_CUDNN_BINVER=\"${CUDNN_MAJOR_VERSION}\"")
math(EXPR CUDNN_VERSION
"${CUDNN_MAJOR_VERSION} * 1000 +
${CUDNN_MINOR_VERSION} * 100 + ${CUDNN_PATCHLEVEL_VERSION}")

@ -32,4 +32,4 @@ endif()
add_dependencies(cub extern_cub)
LIST(APPEND externl_project_dependencies cub)
LIST(APPEND external_project_dependencies cub)

@ -28,4 +28,4 @@ endif()
add_dependencies(dlpack extern_dlpack)
LIST(APPEND externl_project_dependencies dlpack)
LIST(APPEND external_project_dependencies dlpack)

@ -37,13 +37,12 @@ INCLUDE(GNUInstallDirs)
INCLUDE(ExternalProject)
SET(NGRAPH_PROJECT "extern_ngraph")
SET(NGRAPH_VERSION "0.9")
SET(NGRAPH_GIT_TAG "f9fd9d4cc318dc59dd4b68448e7fbb5f67a28bd0")
SET(NGRAPH_GIT_TAG "v0.10.1")
SET(NGRAPH_SOURCES_DIR ${THIRD_PARTY_PATH}/ngraph)
SET(NGRAPH_INSTALL_DIR ${THIRD_PARTY_PATH}/install/ngraph)
SET(NGRAPH_INC_DIR ${NGRAPH_INSTALL_DIR}/include)
SET(NGRAPH_LIB_DIR ${NGRAPH_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR})
SET(NGRAPH_SHARED_LIB_NAME libngraph.so.${NGRAPH_VERSION})
SET(NGRAPH_SHARED_LIB_NAME libngraph.so)
SET(NGRAPH_CPU_LIB_NAME libcpu_backend.so)
SET(NGRAPH_TBB_LIB_NAME libtbb.so.2)
SET(NGRAPH_GIT_REPO "https://github.com/NervanaSystems/ngraph.git")

@ -110,7 +110,7 @@ function(op_library TARGET)
# Define operators that don't need pybind here.
foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op"
"tensor_array_read_write_op" "tensorrt_engine_op" "conv_fusion_op"
"fusion_transpose_flatten_concat_op")
"fusion_transpose_flatten_concat_op" "fusion_conv_inception_op")
if ("${TARGET}" STREQUAL "${manual_pybind_op}")
set(pybind_flag 1)
endif()

@ -27,9 +27,10 @@ add_subdirectory(details)
proto_library(framework_proto SRCS framework.proto)
proto_library(async_executor_proto SRCS data_feed.proto)
cc_library(ddim SRCS ddim.cc DEPS eigen3 boost)
cc_library(ddim SRCS ddim.cc DEPS eigen3 boost enforce)
cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
nv_test(dim_test SRCS dim_test.cu DEPS ddim)
cc_test(unroll_array_ops_test SRCS unroll_array_ops_test.cc)
cc_library(data_type SRCS data_type.cc DEPS framework_proto ddim device_context)
cc_test(data_type_test SRCS data_type_test.cc DEPS data_type place tensor)
if(WITH_GPU)
@ -71,13 +72,13 @@ cc_test(reader_test SRCS reader_test.cc DEPS reader)
cc_library(threadpool SRCS threadpool.cc DEPS enforce)
cc_test(threadpool_test SRCS threadpool_test.cc DEPS threadpool)
cc_library(var_type_traits SRCS var_type_traits DEPS lod_tensor selected_rows framework_proto)
cc_library(var_type_traits SRCS var_type_traits DEPS lod_tensor selected_rows framework_proto)
if (WITH_GPU)
target_link_libraries(var_type_traits dynload_cuda)
endif()
cc_test(var_type_traits_test SRCS var_type_traits_test.cc DEPS var_type_traits)
cc_library(scope SRCS scope.cc DEPS glog threadpool var_type_traits)
cc_library(scope SRCS scope.cc DEPS glog threadpool xxhash var_type_traits)
cc_library(scope_pool SRCS scope_pool.cc DEPS scope)
cc_test(scope_test SRCS scope_test.cc DEPS scope)
cc_test(variable_test SRCS variable_test.cc DEPS tensor var_type_traits)
@ -129,11 +130,9 @@ cc_test(version_test SRCS version_test.cc DEPS version)
cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog version)
if(WITH_NGRAPH)
if(NOT WIN32)
cc_library(ngraph_bridge SRCS ngraph_bridge.cc DEPS operator framework_proto ngraph)
cc_library(ngraph_operator SRCS ngraph_operator.cc DEPS ngraph_bridge operator op_info device_context tensor scope glog
shape_inference data_transform lod_tensor profiler ngraph)
endif(NOT WIN32)
cc_library(ngraph_bridge SRCS ngraph_bridge.cc DEPS operator framework_proto ngraph)
cc_library(ngraph_operator SRCS ngraph_operator.cc DEPS ngraph_bridge operator op_info device_context tensor scope glog
shape_inference data_transform lod_tensor profiler)
endif(WITH_NGRAPH)
cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc)
@ -175,11 +174,7 @@ if(WITH_DISTRIBUTE)
else()
if(WITH_NGRAPH)
if(NOT WIN32)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass ngraph ngraph_operator variable_helper)
else(NOT WIN32)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
endif(NOT WIN32)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass ngraph_operator variable_helper)
else(WITH_NGRAPH)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
endif(WITH_NGRAPH)
@ -194,9 +189,9 @@ cc_library(parallel_executor SRCS parallel_executor.cc DEPS
fast_threaded_ssa_graph_executor variable_helper)
if(WITH_PSLIB)
cc_library(async_executor SRCS async_executor.cc data_feed.cc data_feed_factory.cc executor_thread_worker.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass async_executor_proto variable_helper pslib_brpc pslib)
cc_library(async_executor SRCS async_executor.cc data_feed.cc data_feed_factory.cc executor_thread_worker.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass async_executor_proto variable_helper pslib_brpc pslib timer)
else()
cc_library(async_executor SRCS async_executor.cc data_feed.cc data_feed_factory.cc executor_thread_worker.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass async_executor_proto variable_helper)
cc_library(async_executor SRCS async_executor.cc data_feed.cc data_feed_factory.cc executor_thread_worker.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass async_executor_proto variable_helper timer)
endif(WITH_PSLIB)

@ -15,34 +15,123 @@
#pragma once
#include <cstdint>
#include "paddle/fluid/platform/hostdevice.h"
#include "paddle/fluid/framework/unroll_array_ops.h"
#include "paddle/fluid/platform/enforce.h"
namespace paddle {
namespace framework {
template <typename T, size_t N>
class Array {
static_assert(N > 0, "The size of array must be larger than 0");
public:
HOSTDEVICE Array() {}
static constexpr size_t kSize = N;
HOSTDEVICE inline Array() {}
HOSTDEVICE explicit Array(const T &val) {
for (size_t i = 0; i < N; ++i) data_[i] = val;
template <typename... Args>
HOSTDEVICE inline explicit Array(const T &val, Args... args) {
static_assert(N == sizeof...(Args) + 1, "Invalid argument");
UnrollVarArgsAssign<T>::Run(data_, val, args...);
}
HOSTDEVICE const T *Get() const { return data_; }
HOSTDEVICE inline void Fill(const T &val) {
UnrollFillConstant<N>::Run(data_, val);
}
HOSTDEVICE T *GetMutable() { return data_; }
HOSTDEVICE inline const T *Get() const { return data_; }
HOSTDEVICE T &operator[](size_t index) { return data_[index]; }
HOSTDEVICE inline T *GetMutable() { return data_; }
HOSTDEVICE const T &operator[](size_t index) const { return data_[index]; }
HOSTDEVICE inline T &operator[](size_t i) { return *advance(data_, i); }
// Writing "return data_[i]" would cause compilation warning/error:
// "array subscript is above array bound" in Python 35 CI.
// It seems that it is a false warning of GCC if we do not check the bounds
// of array index. But for better performance, we do not check in operator[]
// like what is in STL. If users want to check the bounds, use at() instead
HOSTDEVICE inline const T &operator[](size_t i) const {
return *advance(data_, i);
}
HOSTDEVICE inline T &at(size_t i) {
#ifndef __CUDA_ARCH__
PADDLE_ENFORCE_LT(i, N, "Array index out of bounds");
#endif
return (*this)[i];
}
HOSTDEVICE inline const T &at(size_t i) const {
#ifndef __CUDA_ARCH__
PADDLE_ENFORCE_LT(i, N, "Array index out of bounds");
#endif
return (*this)[i];
}
HOSTDEVICE constexpr size_t size() const { return N; }
HOSTDEVICE inline bool operator==(const Array<T, N> &other) const {
return UnrollCompare<N>::Run(data_, other.data_);
}
HOSTDEVICE inline bool operator!=(const Array<T, N> &other) const {
return !(*this == other);
}
private:
template <typename U>
HOSTDEVICE static inline U *advance(U *ptr, size_t i) {
return ptr + i;
}
T data_[N];
};
template <typename T>
class Array<T, 0> {
public:
static constexpr size_t kSize = 0;
HOSTDEVICE inline Array() {}
HOSTDEVICE inline void Fill(const T &val) {}
HOSTDEVICE inline constexpr T *Get() const { return nullptr; }
// Add constexpr to GetMutable() cause warning in MAC
HOSTDEVICE inline T *GetMutable() { return nullptr; }
HOSTDEVICE inline T &operator[](size_t) {
#ifdef __CUDA_ARCH__
static T obj();
return obj;
#else
PADDLE_THROW("Array<T, 0> has no element");
#endif
}
HOSTDEVICE inline const T &operator[](size_t) const {
#ifdef __CUDA_ARCH__
static const T obj();
return obj;
#else
PADDLE_THROW("Array<T, 0> has no element");
#endif
}
HOSTDEVICE inline T &at(size_t i) { return (*this)[i]; }
HOSTDEVICE inline const T &at(size_t i) const { return (*this)[i]; }
HOSTDEVICE constexpr size_t size() const { return 0; }
HOSTDEVICE constexpr bool operator==(const Array<T, 0> &other) const {
return true;
}
HOSTDEVICE constexpr bool operator!=(const Array<T, 0> &other) const {
return false;
}
};
} // namespace framework
} // namespace paddle

@ -304,8 +304,13 @@ void AsyncExecutor::RunFromFile(const ProgramDesc& main_program,
// start executing ops in multiple threads
for (int thidx = 0; thidx < actual_thread_num; ++thidx) {
threads.push_back(
std::thread(&ExecutorThreadWorker::TrainFiles, workers[thidx].get()));
if (debug) {
threads.push_back(std::thread(&ExecutorThreadWorker::TrainFilesWithTimer,
workers[thidx].get()));
} else {
threads.push_back(
std::thread(&ExecutorThreadWorker::TrainFiles, workers[thidx].get()));
}
}
for (auto& th : threads) {

File diff suppressed because it is too large Load Diff

@ -18,62 +18,145 @@ limitations under the License. */
#include <stdexcept>
#include <vector>
#include "paddle/fluid/framework/dim.h"
#include "paddle/fluid/platform/enforce.h"
#include "paddle/fluid/platform/variant.h"
namespace paddle {
namespace framework {
#define PADDLE_VISIT_DDIM_BASE(rank, callback) \
case (rank): { \
constexpr auto kRank = (rank); \
return (callback); \
}
#define PADDLE_VISIT_DDIM(rank, callback) \
switch (rank) { \
PADDLE_VISIT_DDIM_BASE(0, callback); \
PADDLE_VISIT_DDIM_BASE(1, callback); \
PADDLE_VISIT_DDIM_BASE(2, callback); \
PADDLE_VISIT_DDIM_BASE(3, callback); \
PADDLE_VISIT_DDIM_BASE(4, callback); \
PADDLE_VISIT_DDIM_BASE(5, callback); \
PADDLE_VISIT_DDIM_BASE(6, callback); \
PADDLE_VISIT_DDIM_BASE(7, callback); \
PADDLE_VISIT_DDIM_BASE(8, callback); \
PADDLE_VISIT_DDIM_BASE(9, callback); \
default: \
PADDLE_THROW("Invalid rank %d", rank); \
}
template <typename T1, typename T2>
inline void dynamic_dim_assign(const T1* in, T2* out, int n) {
PADDLE_VISIT_DDIM(n, (static_dim_assign<kRank, T1, T2>(in, out)));
}
/**
* \brief A dynamically sized dimension.
*
* The number of dimensions must be between [1, 9].
*/
struct DDim {
typedef boost::variant<Dim<0>, Dim<1>, Dim<2>, Dim<3>, Dim<4>, Dim<5>, Dim<6>,
Dim<7>, Dim<8>, Dim<9>>
DDimVar;
DDimVar var;
class DDim {
public:
constexpr static int kMaxRank = 9;
DDim() : rank_(1) { dim_[0] = 0; }
DDim() : var(Dim<1>()) {}
DDim(const DDim& ddim) : dim_() { CopyFrom(ddim); }
DDim(const int* d, int n) : rank_(n) {
dynamic_dim_assign(d, dim_.GetMutable(), n);
}
DDim(const int64_t* d, int n) : rank_(n) {
dynamic_dim_assign(d, dim_.GetMutable(), n);
}
template <int D>
explicit DDim(const Dim<D>& in) : var(in) {}
/*implicit*/ DDim(const Dim<D>& in) : rank_(D) { // NOLINT
UnsafeCast<D>() = in;
}
/*implicit*/ DDim(std::initializer_list<int64_t> init_list)
: DDim(init_list.begin(), init_list.size()) {}
/*implicit*/ DDim(std::initializer_list<int64_t> init_list);
inline DDim& operator=(const DDim& ddim) { return CopyFrom(ddim); }
template <int D>
DDim& operator=(const Dim<D>& in) {
var = in;
inline DDim& operator=(const Dim<D>& dim) {
rank_ = D;
UnsafeCast<D>() = dim;
return *this;
}
int64_t& operator[](int idx);
int64_t operator[](int idx) const;
inline int64_t& operator[](int idx) { return dim_[idx]; }
inline int64_t operator[](int idx) const { return dim_[idx]; }
inline int64_t& at(int idx) {
PADDLE_ENFORCE(idx >= 0 && idx < rank_, "Invalid idx %d", idx);
return dim_[idx];
}
inline int64_t at(int idx) const {
PADDLE_ENFORCE(idx >= 0 && idx < rank_, "Invalid idx %d", idx);
return dim_[idx];
}
template <typename Visitor>
typename Visitor::result_type apply_visitor(Visitor& visitor) {
return var.apply_visitor(visitor);
typename std::result_of<Visitor(Dim<0>&)>::type apply_visitor(
Visitor&& visitor) {
PADDLE_VISIT_DDIM(rank_, visitor(UnsafeCast<kRank>()));
}
template <typename Visitor>
typename Visitor::result_type apply_visitor(Visitor& visitor) const {
return var.apply_visitor(visitor);
typename std::result_of<Visitor(const Dim<0>&)>::type apply_visitor(
Visitor&& visitor) const {
PADDLE_VISIT_DDIM(rank_, visitor(UnsafeCast<kRank>()));
}
DDimVar getVar() { return var; }
bool operator==(const DDim& d) const;
bool operator!=(const DDim& d) const;
DDim operator+(const DDim& d) const;
bool operator==(DDim d) const;
DDim operator*(const DDim& d) const;
bool operator!=(DDim d) const;
inline const int64_t* Get() const { return dim_.Get(); }
DDim operator+(DDim d) const;
inline int64_t* GetMutable() { return dim_.GetMutable(); }
DDim operator*(DDim d) const;
inline int size() const { return rank_; }
private:
template <int D>
inline Dim<D>& UnsafeCast() {
static_assert(D >= 0 && D <= kMaxRank, "Invalid rank");
auto* p = static_cast<void*>(&dim_);
return *reinterpret_cast<Dim<D>*>(p);
}
template <int D>
inline const Dim<D>& UnsafeCast() const {
static_assert(D >= 0 && D <= kMaxRank, "Invalid rank");
auto* p = static_cast<const void*>(&dim_);
return *reinterpret_cast<const Dim<D>*>(p);
}
int size() const;
inline DDim& CopyFrom(const DDim& ddim) {
PADDLE_VISIT_DDIM(ddim.rank_, (*this = ddim.UnsafeCast<kRank>()));
}
friend DDim stride(const DDim& ddim);
friend DDim stride_numel(const DDim& ddim);
private:
Dim<kMaxRank> dim_;
int rank_;
};
#undef PADDLE_VISIT_DDIM_BASE
#undef PADDLE_VISIT_DDIM
/**
* \brief Make a DDim from std::vector<int64_t>
*
@ -92,7 +175,7 @@ DDim make_ddim(const std::vector<int>& dims);
DDim make_ddim(std::initializer_list<int64_t> dims);
int64_t get(const DDim& dim, int idx);
void set(DDim& dim, int idx, int val);
void set(DDim& dim, int idx, int val); // NOLINT
std::vector<int64_t> vectorize(const DDim& ddim);
std::vector<int> vectorize2int(const DDim& ddim);
@ -129,12 +212,3 @@ DDim stride(const DDim& ddim);
DDim stride_numel(const DDim& ddim);
} // namespace framework
} // namespace paddle
namespace boost {
template <typename T>
T get(const paddle::framework::DDim& in) {
return boost::get<T>(in.var);
}
} // namespace boost

@ -50,7 +50,7 @@ void AllReduceOpHandle::RunImpl() {
// FIXME(typhoonzero): If scope0(global scope) have NCCL_ID_VAR,
// this is a distributed or inter-process call, find a better way.
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
if (NoDummyInputSize() == 1 &&
local_scopes_[0]->FindLocalVar(NCCL_ID_VARNAME) == nullptr) {
#else

@ -25,7 +25,7 @@ struct ExecutionStrategy {
size_t num_threads_{0};
bool use_cuda_{true};
bool allow_op_delay_{false};
size_t num_iteration_per_drop_scope_{100};
size_t num_iteration_per_drop_scope_{1};
ExecutorType type_{kDefault};
bool dry_run_{false};
};

@ -64,20 +64,26 @@ FeedFetchList ScopeBufferedSSAGraphExecutor::Run(
}
platform::RecordEvent e("ScopeBufferedSSAGraphExecutorAfterRun", nullptr);
drop_scope_counter_ += 1;
++drop_scope_counter_;
if (!fetch_tensors.empty() ||
drop_scope_counter_ == strategy_.num_iteration_per_drop_scope_) {
drop_scope_counter_ = 0;
// Wait All computational streams
for (auto p : places_) {
platform::DeviceContextPool::Instance().Get(p)->Wait();
bool stream_end = false;
if (!fetch_tensors.empty()) {
WaitComputationalStreams();
stream_end = true;
}
if (drop_scope_counter_ == strategy_.num_iteration_per_drop_scope_) {
if (!stream_end) {
WaitComputationalStreams();
}
for (auto &scope : local_scopes_) {
auto &local_scope =
*scope->Var(details::kLocalExecScopeName)->GetMutable<Scope *>();
scope->DeleteScope(local_scope);
}
drop_scope_counter_ = 0;
}
if (eptr) {
std::rethrow_exception(eptr);

@ -47,6 +47,14 @@ class ScopeBufferedSSAGraphExecutor : public SSAGraphExecutor {
FeedFetchList Run(const std::vector<std::string>& fetch_tensors) override;
private:
inline void WaitComputationalStreams() {
// Wait All computational streams
for (auto p : places_) {
platform::DeviceContextPool::Instance().Get(p)->Wait();
}
}
private:
size_t drop_scope_counter_{0};

File diff suppressed because it is too large Load Diff

@ -59,7 +59,7 @@ static DLDataType GetDLDataTypeFromTypeIndex(proto::VarType::Type type) {
struct DLContextVisitor : public boost::static_visitor<::DLContext> {
inline ::DLContext operator()(const platform::CPUPlace &place) const {
DLContext ctx;
::DLContext ctx;
ctx.device_type = kDLCPU;
ctx.device_id = 0;
return ctx;
@ -67,7 +67,7 @@ struct DLContextVisitor : public boost::static_visitor<::DLContext> {
inline ::DLContext operator()(const platform::CUDAPlace &place) const {
#ifdef PADDLE_WITH_CUDA
DLContext ctx;
::DLContext ctx;
ctx.device_type = kDLGPU;
ctx.device_id = place.device;
return ctx;
@ -78,7 +78,7 @@ struct DLContextVisitor : public boost::static_visitor<::DLContext> {
inline ::DLContext operator()(const platform::CUDAPinnedPlace &place) const {
#ifdef PADDLE_WITH_CUDA
DLContext ctx;
::DLContext ctx;
ctx.device_type = kDLCPUPinned;
ctx.device_id = 0;
return ctx;

@ -38,7 +38,7 @@ class DLPackTensor {
// The shape in DLTensor is defined as int64_t*
// Add this member to make TVMTensor init without heap allocation
ShapeType shape_[9];
ShapeType shape_[DDim::kMaxRank];
};
} // namespace framework

@ -22,7 +22,7 @@ limitations under the License. */
#include "paddle/fluid/framework/reader.h"
#include "paddle/fluid/framework/transfer_scope_cache.h"
#include "paddle/fluid/framework/variable_helper.h"
#include "paddle/fluid/operators/detail/macros.h"
#include "paddle/fluid/operators/distributed/distributed.h"
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/platform/profiler.h"

@ -29,6 +29,7 @@ limitations under the License. */
#include "paddle/fluid/inference/io.h"
#include "paddle/fluid/platform/cpu_helper.h"
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/platform/timer.h"
#include "paddle/fluid/pybind/pybind.h"
namespace paddle {
namespace framework {
@ -180,6 +181,7 @@ void ExecutorThreadWorker::SetDevice() {
return;
#else
static unsigned concurrency_cap = std::thread::hardware_concurrency();
LOG(WARNING) << "concurrency capacity " << concurrency_cap;
int thread_id = this->thread_id_;
if (static_cast<unsigned>(thread_id) < concurrency_cap) {
@ -238,6 +240,55 @@ static void print_fetch_var(Scope* scope, const std::string& var_name) {
VLOG(1) << "print_fetch_var: unrecognized data type:" << tensor.type();
}
void ExecutorThreadWorker::TrainFilesWithTimer() {
platform::SetNumThreads(1);
SetDevice();
thread_reader_->Start();
std::vector<double> op_total_time;
std::vector<std::string> op_name;
for (auto& op : ops_) {
op_name.push_back(op->Type());
}
op_total_time.resize(ops_.size());
for (size_t i = 0; i < op_total_time.size(); ++i) {
op_total_time[i] = 0.0;
}
platform::Timer timeline;
double total_time = 0.0;
double read_time = 0.0;
int cur_batch;
int batch_cnt = 0;
timeline.Start();
while ((cur_batch = thread_reader_->Next()) > 0) {
timeline.Pause();
read_time += timeline.ElapsedSec();
total_time += timeline.ElapsedSec();
for (size_t i = 0; i < ops_.size(); ++i) {
timeline.Start();
ops_[i]->Run(*thread_scope_, place_);
timeline.Pause();
op_total_time[i] += timeline.ElapsedSec();
total_time += timeline.ElapsedSec();
}
++batch_cnt;
thread_scope_->DropKids();
if (thread_id_ == 0) {
if (batch_cnt > 0 && batch_cnt % 1000 == 0) {
for (size_t i = 0; i < ops_.size(); ++i) {
fprintf(stderr, "op_name:[%zu][%s], op_mean_time:[%fs]\n", i,
op_name[i].c_str(), op_total_time[i] / batch_cnt);
}
fprintf(stderr, "mean read time: %fs\n", read_time / batch_cnt);
int fetch_var_num = fetch_var_names_.size();
for (int i = 0; i < fetch_var_num; ++i) {
print_fetch_var(thread_scope_, fetch_var_names_[i]);
}
}
}
timeline.Start();
}
}
void ExecutorThreadWorker::TrainFiles() {
platform::SetNumThreads(1);
@ -320,10 +371,12 @@ void AsyncExecutorThreadWorker::SetPSlibPtr(
std::shared_ptr<paddle::distributed::PSlib> pslib_ptr) {
_pslib_ptr = pslib_ptr;
}
void AsyncExecutorThreadWorker::SetPullDenseThread(
std::shared_ptr<DensePullThread> dpt) {
_pull_dense_thread = dpt;
}
void AsyncExecutorThreadWorker::TrainOneNetwork() {
PrepareParams();

@ -155,6 +155,8 @@ class ExecutorThreadWorker {
void SetDataFeed(const std::shared_ptr<DataFeed>& datafeed);
// A multi-thread training function
virtual void TrainFiles();
// with timer log
virtual void TrainFilesWithTimer();
// set fetch variable names from python interface assigned by users
void SetFetchVarNames(const std::vector<std::string>& fetch_var_names);
#ifdef PADDLE_WITH_PSLIB

@ -399,7 +399,7 @@ void NgraphEngine::BuildNgFunction() {
BuildNgNodes();
ngraph_function_ = nullptr;
ngraph::NodeVector func_outputs;
ngraph::op::ParameterVector func_inputs;
ngraph::ParameterVector func_inputs;
for (auto& vo : var_out_) {
func_outputs.push_back(var_node_map_->at(vo));

@ -16,7 +16,9 @@ limitations under the License. */
#if !defined(_WIN32)
#include <pthread.h>
#endif // !_WIN32
#else
#include <mutex> // NOLINT
#endif // !_WIN32
#include "paddle/fluid/platform/enforce.h"
@ -29,17 +31,17 @@ struct RWLock {
~RWLock() { pthread_rwlock_destroy(&lock_); }
void RDLock() {
inline void RDLock() {
PADDLE_ENFORCE_EQ(pthread_rwlock_rdlock(&lock_), 0,
"acquire read lock failed");
}
void WRLock() {
inline void WRLock() {
PADDLE_ENFORCE_EQ(pthread_rwlock_wrlock(&lock_), 0,
"acquire write lock failed");
}
void UNLock() {
inline void UNLock() {
PADDLE_ENFORCE_EQ(pthread_rwlock_unlock(&lock_), 0, "unlock failed");
}
@ -51,81 +53,46 @@ struct RWLock {
// https://stackoverflow.com/questions/7125250/making-pthread-rwlock-wrlock-recursive
// In windows, rw_lock seems like a hack. Use empty object and do nothing.
struct RWLock {
void RDLock() {}
void WRLock() {}
void UNLock() {}
// FIXME(minqiyang): use mutex here to do fake lock
inline void RDLock() { mutex_.lock(); }
inline void WRLock() { mutex_.lock(); }
inline void UNLock() { mutex_.unlock(); }
private:
std::mutex mutex_;
};
#endif
class RWLockGuard {
class AutoWRLock {
public:
enum Status { kUnLock, kWRLock, kRDLock };
RWLockGuard(RWLock* rw_lock, Status init_status)
: lock_(rw_lock), status_(Status::kUnLock) {
switch (init_status) {
case Status::kRDLock: {
RDLock();
break;
}
case Status::kWRLock: {
WRLock();
break;
}
case Status::kUnLock: {
break;
}
}
}
explicit AutoWRLock(RWLock* rw_lock) : lock_(rw_lock) { Lock(); }
void WRLock() {
switch (status_) {
case Status::kUnLock: {
lock_->WRLock();
status_ = Status::kWRLock;
break;
}
case Status::kWRLock: {
break;
}
case Status::kRDLock: {
PADDLE_THROW(
"Please unlock read lock first before invoking write lock.");
break;
}
}
}
~AutoWRLock() { UnLock(); }
void RDLock() {
switch (status_) {
case Status::kUnLock: {
lock_->RDLock();
status_ = Status::kRDLock;
break;
}
case Status::kRDLock: {
break;
}
case Status::kWRLock: {
PADDLE_THROW(
"Please unlock write lock first before invoking read lock.");
break;
}
}
}
private:
inline void Lock() { lock_->WRLock(); }
void UnLock() {
if (status_ != Status::kUnLock) {
lock_->UNLock();
status_ = Status::kUnLock;
}
}
inline void UnLock() { lock_->UNLock(); }
private:
RWLock* lock_;
};
class AutoRDLock {
public:
explicit AutoRDLock(RWLock* rw_lock) : lock_(rw_lock) { Lock(); }
~AutoRDLock() { UnLock(); }
private:
inline void Lock() { lock_->RDLock(); }
~RWLockGuard() { UnLock(); }
inline void UnLock() { lock_->UNLock(); }
private:
RWLock* lock_;
Status status_;
};
} // namespace framework

@ -47,9 +47,15 @@ DEFINE_bool(fast_eager_deletion_mode, false,
// the mutex will cause serious performance issue.
// So the mutex is disabled when `ON_INFER`.
#ifdef PADDLE_ON_INFERENCE
#define SCOPE_LOCK_GUARD
#define SCOPE_KIDS_READER_LOCK
#define SCOPE_KIDS_WRITER_LOCK
#define SCOPE_VARS_READER_LOCK
#define SCOPE_VARS_WRITER_LOCK
#else
#define SCOPE_LOCK_GUARD std::lock_guard<std::mutex> lock(mutex_);
#define SCOPE_KIDS_READER_LOCK AutoRDLock auto_lock(&kids_lock_);
#define SCOPE_KIDS_WRITER_LOCK AutoWRLock auto_lock(&kids_lock_);
#define SCOPE_VARS_READER_LOCK AutoRDLock auto_lock(&vars_lock_);
#define SCOPE_VARS_WRITER_LOCK AutoWRLock auto_lock(&vars_lock_);
#endif
namespace paddle {
@ -67,64 +73,69 @@ bool IsFastEagerDeletionModeEnabled() { return FLAGS_fast_eager_deletion_mode; }
Scope::~Scope() { DropKids(); }
Scope& Scope::NewScope() const {
SCOPE_LOCK_GUARD
kids_.push_back(new Scope(this));
return *kids_.back();
Scope* child = new Scope(this);
{
SCOPE_KIDS_WRITER_LOCK
kids_.push_back(child);
}
return *child;
}
Variable* Scope::Var(const std::string& name) {
SCOPE_LOCK_GUARD
SCOPE_VARS_WRITER_LOCK
return VarInternal(name);
}
Variable* Scope::Var(std::string* name) {
SCOPE_LOCK_GUARD
auto new_name = string::Sprintf("%p.%d", this, vars_.size());
if (name != nullptr) {
*name = new_name;
}
SCOPE_VARS_WRITER_LOCK
return VarInternal(new_name);
}
Variable* Scope::FindVar(const std::string& name) const {
SCOPE_LOCK_GUARD
SCOPE_VARS_READER_LOCK
return FindVarInternal(name);
}
Variable* Scope::FindLocalVar(const std::string& name) const {
SCOPE_LOCK_GUARD
SCOPE_VARS_READER_LOCK
return FindVarLocally(name);
}
const Scope* Scope::FindScope(const Variable* var) const {
SCOPE_LOCK_GUARD
SCOPE_VARS_READER_LOCK
return FindScopeInternal(var);
}
void Scope::DropKids() {
SCOPE_LOCK_GUARD
SCOPE_KIDS_WRITER_LOCK
for (Scope* s : kids_) delete s;
kids_.clear();
}
bool Scope::HasKid(const Scope* scope) const {
SCOPE_LOCK_GUARD
SCOPE_KIDS_READER_LOCK
auto it = std::find(this->kids_.begin(), this->kids_.end(), scope);
return it != this->kids_.end();
}
std::vector<std::string> Scope::LocalVarNames() const {
SCOPE_LOCK_GUARD
std::vector<std::string> known_vars;
known_vars.reserve(this->vars_.size());
for (auto& p : vars_) {
known_vars.emplace_back(p.first);
{
SCOPE_VARS_READER_LOCK
known_vars.reserve(this->vars_.size());
for (auto& p : vars_) {
known_vars.emplace_back(p.first);
}
}
return known_vars;
}
void Scope::DeleteScope(Scope* scope) const {
SCOPE_LOCK_GUARD
SCOPE_KIDS_WRITER_LOCK
auto it = std::find(this->kids_.begin(), this->kids_.end(), scope);
PADDLE_ENFORCE(it != this->kids_.end(), "%p Cannot find %p as kid scope",
this, scope);
@ -138,8 +149,8 @@ void Scope::DeleteScope(Scope* scope) const {
}
void Scope::EraseVars(const std::vector<std::string>& var_names) {
SCOPE_LOCK_GUARD
std::set<std::string> var_set(var_names.begin(), var_names.end());
SCOPE_VARS_WRITER_LOCK
for (auto it = vars_.begin(); it != vars_.end();) {
if (var_set.find(it->first) != var_set.end()) {
it = vars_.erase(it);
@ -151,12 +162,12 @@ void Scope::EraseVars(const std::vector<std::string>& var_names) {
void Scope::Rename(const std::string& origin_name,
const std::string& new_name) const {
SCOPE_LOCK_GUARD
SCOPE_VARS_WRITER_LOCK
RenameInternal(origin_name, new_name);
}
std::string Scope::Rename(const std::string& origin_name) const {
SCOPE_LOCK_GUARD
SCOPE_VARS_WRITER_LOCK
auto new_name = string::Sprintf("%p.%d", this, vars_.size());
RenameInternal(origin_name, new_name);
return new_name;

@ -14,12 +14,18 @@ limitations under the License. */
#pragma once
extern "C" {
#include <xxhash.h>
}
#include <list>
#include <mutex> // NOLINT
#include <memory>
#include <string>
#include <unordered_map>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/rw_lock.h"
#include "paddle/fluid/framework/variable.h"
#include "paddle/fluid/platform/macros.h"
@ -95,7 +101,14 @@ class Scope {
std::string Rename(const std::string& origin_name) const;
protected:
mutable std::unordered_map<std::string, std::unique_ptr<Variable>> vars_;
struct KeyHasher {
std::size_t operator()(const std::string& key) const {
return XXH32(key.c_str(), key.size(), 1);
}
};
mutable std::unordered_map<std::string, std::unique_ptr<Variable>, KeyHasher>
vars_;
private:
// Call Scope::NewScope for a sub-scope.
@ -124,7 +137,8 @@ class Scope {
DISABLE_COPY_AND_ASSIGN(Scope);
private:
mutable std::mutex mutex_;
mutable RWLock kids_lock_;
mutable RWLock vars_lock_;
};
// Generate some debug string about the inherience structure of scope, quite

Some files were not shown because too many files have changed in this diff Show More

Loading…
Cancel
Save